Project Heimdallr
For this challenge, we invite you to become "virtual contributors" to the Asteroid Grand Challenge and develop a hypothetical method, concept note or simple prototype that demonstrates how Machine Learning could be used to help us avoid the same fate as the dinosaurs.
Like the Norse God Heimdallr watching for the onset of Ragnarok, our solution strives to be a guardian for all mankind, helping us avoid the same fate as dinosaurs.
We know that most of the 1 Km radius asteroids that are near Earth objects have been discovered but we also know that only about 50% of the NEOs with the radius between 500m and 1Km have been discovered. The importance of discovering and then tracking all these potentially hazardous objects cannot be stressed enough because the consequences of such an impact can be disastrous an potentially civilization ending. We all remember the Chelyabinsk meteor that fell near the Chelyabinsk town in February 2013 which caused injuries to about 1500 people and this was only a 20 m across piece of rock.
Because we believe that discovery and tracking of NEO's is very important, we thought about how Machine Learning can be used to help with this process, making it more automated and accurate. What we want to propose is a concept of how machine learning can be used to aid with discovery and tracking of asteroids.
Since the time was short and our knowledge of Machine Learning is somewhat lacking - we wanted to learn about that too - we did not manage to implement a working prototype for the purposes of the challenge so we resumed ourselves to providing a solid concept and, following the competition, we can continue working on our software solution for predicting orbits based on astrometric data observations using TensofFlow and deep learning.We strongly believe that since deep learning is picking up such momentum these days the potential of using it for asteroid detection and orbit prediction is very big.
Below we want to give some details about our concept and how ML can help with NEO detection.
The process of detecting NEO's is very complex and it consists of several steps of which here are two relevant for applying machine learning:
- Calculating the observation parameters - this is done by professional and amateur astronomers and is based on taking multiple images of a piece of the sky and then trying to determine if there are moving objects in them. During this step, machine learning can help with detecting if the out of place objects in the observation images are asteroids or if they are lens dirt or cosmic rays or other non-relevant data. This part would greatly help with the reduction phase in which the astronomer obtains the astrometric observation data for the objects and which it is crucial to be quick and accurate. Otherwise, if it takes to long, the asteroid can move out of the field of view and be lost for subsequent observations and tracking.
- Another relevant step in NEO discovery and tracking for machine learning is the calculation of orbital data based on observation data. We believe that deep learning can be used to predict the orbits of NEO's based on observations. Currently there is a load of observation data available for all known NEOs as well as orbital data for them. Feeding all this data into a deep learning net and training it, should allow for prediction of orbits of NEOs which can be used to determine if we are dealing with an already existing object or a new one. This is also a time consuming part since it requires a lot of calculations to get from the observation data to the orbit data and confirm the NEO.
More than this, automating most of the discovery and tracking process, has the potential to create a system which uses robotic telescopes to automatically track and discover NEO's and be backed by a global system in which astronomers can be alerted when something is amiss in the system in order to be able to quickly observe and confirm/infirm the status of a given object. One of the big problems we encountered while working on a prototype is that the MPC data seems to be spread across multiple files which makes it harder to compile and use it for ML purposes. A Big Data system would greatly improve the collection of data and would be very useful for mining with ML algorithms.
https://en.wikipedia.org/wiki/N-body_problem
https://en.wikipedia.org/wiki/Chelyabinsk_meteor
https://en.wikipedia.org/wiki/Chicxulub_impactor
http://neo.jpl.nasa.gov/risk/doc/palermo.html
http://neo.jpl.nasa.gov/torino_scale.html
https://en.wikipedia.org/wiki/Palermo_Technical_Im...
